S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Program in Bioengineering (협동과정-바이오엔지니어링전공) Theses (Master's Degree_협동과정-바이오엔지니어링전공)
Sleep Classification using HRV parameters based on Recurrent Neural Network : 깊은 인공신경망을 기반으로 심박변이율 지표를 이용한 수면 단계분석
- 공과대학 협동과정 바이오엔지니어링전공
- Issue Date
- 서울대학교 대학원
- 학위논문 (석사)-- 서울대학교 대학원 : 협동과정 바이오엔지니어링전공, 2016. 8. 박광석.
- This study proposes a method of sleep stage classification using heart rate variability(HRV) and actigraphy features. ECG signals by heart rate monitor and actigraphy signal were measured during polysomnography (PSG). Total 19 features were inputted to artificial neural network classifier or recurrent neural network classifier. For training the model, the procedure finding optimal parameters which make the model accurately predict the sleep stage was necessary and select the parameters resulting in best performance.
In this study, sleep stage was classified into 4 class, wake, rapid eye movement (REM) sleep, slow wave sleep (SWS) and light sleep. Many researches classified sleep stage into 3 class, wake, REM sleep, and NREM sleep, however detection of SWS has important roles in sleep physiology that reflect the recovery of tired brain and memory consolidation.
In the detection of wake stage, three method, sleep onset detection, long term wake detection and short wake detection, were performed. The result of one method was integrated with the result of other method. The method classified the wake stage with average sensitivity 51% specificity 92% Cohens kappa 0.51 accuracy 85%.
In the detection of REM sleep and SWS, 5 models, linear discriminant analysis (LDA), k-nearest neighbor (kNN), support vector mahine (SVM), artificial neural network(ANN) and recurrent neural network(RNN) were applied as classifier. RNN classified the SWS stage with average sensitivity 64.7% specificity 97.4% Cohens kappa 0.615 accuracy 89.8% and the REM stage with average sensitivity 67% specificity 97% Cohens kappa 0.60 , accuracy 91%.
For 4 stage classification, RNN classified the sleep stage with accuracy 71%, and kappa 0.52. The performance in each sleep stage detection and 4 stage classification were estimated by applying subjects in testing set.